PedroMarinhoDev/Moss-Image-Edit-ComfyUI

Moss-Image-Edit LoRA for ComfyUI

chinoll/Moss-Image-Edit, the image-editing LoRA for MOSS-Image, with its keys renamed for ComfyUI's native MOSS-Image model. Edit an image (or combine up to four) with a text instruction.

"Paint the bicycle bright red, keeping the ivy, the brick wall and the lighting the same." Input; the same graph with the LoRA at 0; with the LoRA.

"Attach the wicker basket from image 2 to the front of the bicycle in image 1, mounted on the handlebars, keeping the ivy, the brick wall and the lighting the same."

Use

The LoRA needs the reference images to go through the model in a way stock ComfyUI doesn't do yet, so install the node pack ComfyUI-Moss-Image-Edit-Native. It works with the stock loaders, LoraLoaderModelOnly, samplers and memory management (no diffusers or transformers), and ships the workflows (Templates → ComfyUI-Moss-Image-Edit-Native): one and two references, each for SFT and for Base.

folder	file
loras	moss_image_edit_lora_bf16.safetensors (this repo)
diffusion_models	moss_image_sft_bf16 or moss_image_sft_int8_convrot from Comfy-Org/MOSS-Image, or moss_image_base_int8_convrot / moss_image_base_bf16 (see below)
text_encoders	qwen3.5_9b_w4a8_gptq from Comfy-Org/MOSS-Image, or a bf16 Qwen3.5-9B
vae	flux2-vae_small from Comfy-Org/MOSS-Image, or flux2-vae (same encoder)

hf download PedroMarinhoDev/Moss-Image-Edit-ComfyUI moss_image_edit_lora_bf16.safetensors --local-dir ComfyUI/models/loras

Settings from the original: references at most 512 px on the long side, 30 steps, cfg 4, shift 5 (the model's default), empty negative; euler + simple reproduces the official schedule. LoRA strength 1.0.

The node pack's workflows default to strength 0.25 with 1024 px references instead. That works for single-image edits, though the result follows the input less closely; two-image edits need strength 1.0 (at 0.25 the output is just image 2).

Base or SFT

The LoRA was trained on SFT and also loads on Base. On moss_image_base_int8_convrot, single-image edits are as clean as on SFT (middle). Combining references is weaker on Base: every seed tried left a ghost of image 2 where it sits in its own frame (right, the cleanest of seven seeds). SFT does this much less, so use SFT for multi-reference edits.

What changed

Only the key names; the tensors are the original bf16 values.

	original adapter_model.safetensors	this file
keys	layers.N.proj.X.lora_A / .lora_B	diffusion_model.layers.N.proj.X.lora_A.weight / .lora_B.weight / .alpha
scale	alpha / rank from adapter_config.json	alpha = 64 per module (rank 64, scale 1.0)
targets	q, k, v, o, ffn_up_gate, ffn_down of all 42 blocks (252 linears)	same; ComfyUI maps q/k/v onto the fused qkv of the int8 files
sha256	17a48b8c605dcffdedb92a2b887d18cf316505be90802717669fb6556098dcdf (revision e19b6ab)	f4bf2905dd40ae490ffda37d6be717f3361a4a723225cf0bf1f3b941eb03315d

Made with tools/convert_lora.py from the node pack, which checks the checksum, targets, shapes and rank/alpha. The original adapter config is kept in the file's metadata (adapter_config).

Credits and license

The LoRA was trained and released by chinoll / Chenkin-x (Moss-Image-Edit, released under GPLv3) on MOSS-Image by OpenMOSS (Apache-2.0). The original model repo names no separate license for the weights; this copy follows the project's GPL-3.0. All credit for the training goes to the original authors.